[Openmp-commits] [PATCH] D60578: [OPENMP][NVPTX]Fix dynamic scheduling in L2+ SPMD parallel regions.

[Alexey Bataev via Phabricator via Openmp-commits]

ABataev created this revision.
ABataev added reviewers: gtbercea, kkwli0, grokos.
Herald added subscribers: jdoerfert, guansong.
Herald added a project: OpenMP.

If the kernel is executed in SPMD mode and the L2+ parallel for region
with the dynamic scheduling is executed, dynamic scheduling functions
are called. They expect full runtime support, but SPMD kernels may be
executed without the full runtime. It leads to the runtime crash of the
compiled program. Patch fixes this problem + fixes handling of the
parallelism level in SPMD mode, which is required as part of this patch.


Repository:
  rOMP OpenMP

https://reviews.llvm.org/D60578

Files:
  libomptarget/deviceRTLs/nvptx/src/libcall.cu
  libomptarget/deviceRTLs/nvptx/src/loop.cu
  libomptarget/deviceRTLs/nvptx/src/parallel.cu


Index: libomptarget/deviceRTLs/nvptx/src/parallel.cu
===================================================================
--- libomptarget/deviceRTLs/nvptx/src/parallel.cu
+++ libomptarget/deviceRTLs/nvptx/src/parallel.cu
@@ -407,7 +407,7 @@
   if (checkRuntimeUninitialized(loc)) {
     ASSERT0(LT_FUSSY, checkSPMDMode(loc),
             "Expected SPMD mode with uninitialized runtime.");
-    return parallelLevel;
+    return parallelLevel + 1;
   }
 
   int threadId = GetLogicalThreadIdInBlock(checkSPMDMode(loc));
Index: libomptarget/deviceRTLs/nvptx/src/loop.cu
===================================================================
--- libomptarget/deviceRTLs/nvptx/src/loop.cu
+++ libomptarget/deviceRTLs/nvptx/src/loop.cu
@@ -205,8 +205,12 @@
   INLINE static void dispatch_init(kmp_Ident *loc, int32_t threadId,
                                    kmp_sched_t schedule, T lb, T ub, ST st,
                                    ST chunk) {
-    ASSERT0(LT_FUSSY, checkRuntimeInitialized(loc),
-            "Expected non-SPMD mode + initialized runtime.");
+    if (checkRuntimeUninitialized(loc)) {
+      // In SPMD mode no need to check parallelism level - dynamic scheduling
+      // may appear only in L2 parallel regions with lightweight runtime.
+      ASSERT0(LT_FUSSY, checkSPMDMode(loc), "Expected non-SPMD mode.");
+      return;
+    }
     int tid = GetLogicalThreadIdInBlock(checkSPMDMode(loc));
     omptarget_nvptx_TaskDescr *currTaskDescr = getMyTopTaskDescriptor(tid);
     T tnum = currTaskDescr->ThreadsInTeam();
@@ -439,8 +443,15 @@
 
   INLINE static int dispatch_next(kmp_Ident *loc, int32_t gtid, int32_t *plast,
                                   T *plower, T *pupper, ST *pstride) {
-    ASSERT0(LT_FUSSY, checkRuntimeInitialized(loc),
-            "Expected non-SPMD mode + initialized runtime.");
+    if (checkRuntimeUninitialized(loc)) {
+      // In SPMD mode no need to check parallelism level - dynamic scheduling
+      // may appear only in L2 parallel regions with lightweight runtime.
+      ASSERT0(LT_FUSSY, checkSPMDMode(loc), "Expected non-SPMD mode.");
+      if (*plast)
+        return DISPATCH_FINISHED;
+      *plast = 1;
+      return DISPATCH_NOTFINISHED;
+    }
     // ID of a thread in its own warp
 
     // automatically selects thread or warp ID based on selected implementation
Index: libomptarget/deviceRTLs/nvptx/src/libcall.cu
===================================================================
--- libomptarget/deviceRTLs/nvptx/src/libcall.cu
+++ libomptarget/deviceRTLs/nvptx/src/libcall.cu
@@ -164,7 +164,8 @@
   if (isRuntimeUninitialized()) {
     ASSERT0(LT_FUSSY, isSPMDMode(),
             "Expected SPMD mode only with uninitialized runtime.");
-    return parallelLevel;
+    // parallelLevel starts from 0, need to add 1 for correct level.
+    return parallelLevel + 1;
   }
   int level = 0;
   omptarget_nvptx_TaskDescr *currTaskDescr =


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D60578.194744.patch
Type: text/x-patch
Size: 2913 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/openmp-commits/attachments/20190411/428cb20d/attachment.bin>